Sensitivity Analysis: Leave-One-Out Evaluation

The surrogacy measures estimated in the previous section are derived using all units in the study (i.e., all trials, centers, etc.). To check the stability of the estimated measures, we propose to use a “leave-one-out” evaluation procedure as a sensitivity analysis approach. It is an iterative procedure in which at each iteration, one trial is left out and the surrogacy measures are estimated using the remaining trials in the data. The procedure is shown in Figure 12.2. At the end of the run, parameter estimates for both R^rial and R?ndiv are obtained for each trial and influential trials can be identified.


Age-Related Macular Degeneration Trial. Descriptive plots. Left panel: Patient distribution in the trials by treatment arm. Right panel: Scatter plot between the true and the surrogate endpoints.


The SAS macro %CONTCONTFULL can be used to conduct the analysis for a

setting with two normally distributed endpoints using the full fixed-effects

model. The following macro call is used:

%CONTCONTFULL(data=armd,true=dif f 52,surrog=dif f 24,trt=treat, trial=center,patid=patientId,weighted=1, looa=1)

The optional arguments that we use are:

  • • weighted: an option which allows the use of weighted regression (weighted=1) in the computation of the trial-level surrogacy. The number of patients in the trial is used as the weight.
  • • looa: an option to perform a leave-one-trial-out analysis (looa=1). Both surrogacy measures are computed by leaving each one of the trials out, so as to check the influence of the trial on the surrogacy measures.
< Prev   CONTENTS   Source   Next >